Features

Hand-written autograd

Every differentiable operation wraps its inputs in a dedicated *Backward node that implements the local gradient rule. The full reverse-mode chain is explicit and readable — backpropagation is a simple walk over the graph each op builds.

• Tensor with requires_grad, .backward(), .grad, and PyTorch-style operators: +, @, *, indexing, sin, cos, exp, log, sqrt, tanh, sigmoid, relu, softmax, mean, max, min, norm, abs, and more.
• Non-scalar outputs require an explicit gradient argument to backward().

Neural networks

• nn.Linear, nn.Sequential, activations (ReLU, Tanh, Sigmoid, Softmax, LeakyReLU, …) and losses (MSELoss, CrossEntropyLoss, BCELoss, L1Loss).
• state_dict / save / load for checkpointing — keys use 0.weight-style prefixes for sub-modules.
• Functional helpers: jacobian, vectorize_parameters, one_hot.

14 optimizers

SGD, SGD with momentum, Adam, Adamax, NAdam, AMSGrad, RMSprop, Adagrad, Adadelta, Nesterov, PID, Levenberg–Marquardt, and more — all sharing a single Optimizer base with a standard step() / zero_grad() interface.

Control-flavoured tools

• KalmanFilter — linear-Gaussian state estimation.
• EKFOptimizer — extended-Kalman neural identification of nonlinear plants.
• RLS — recursive least squares for online system identification.
• Hand-rolled jacobian for exact gradients of network outputs w.r.t. parameters.

Verified by finite differences

The test suite checks the gradient of every operation against numeric differentiation, so the math is not just claimed — it is measured. The suite currently has 108 passing tests covering tensors, modules, optimizers, and the control tools.

Quality

• Linting and formatting via ruff (clean).
• CI on GitHub Actions across Python 3.10–3.13.
• MIT licensed.